From Tools to Teammates

Abstract

For seventy years we have used computers as tools. You issue a command, the machine obeys, and the responsibility stays entirely with you. AI agents break that frame. An agent does not just execute, it drafts, chooses, negotiates, and acts on your behalf, which makes it less like a tool and more like a colleague. The defining question of the next decade is not how powerful agents get. It is how humans and agents work together: who owns the work, who is accountable for it, and whether the arrangement leaves people more sovereign or less. This paper argues the tool era is ending, names the hard problems of treating an agent as a participant rather than a device (accountability, trust, ownership, coordination, and the quiet erosion of human agency), and describes the substrate that collaboration needs to stay accountable and human. Surfaces is our first attempt to build it.

The Why: The Tool Era Is Ending

A tool has a clean relationship with you: command, obey, and you are responsible for the result. A hammer has no judgment. A calculator has no stake in the answer. That clarity is the whole reason software has felt safe: whatever it did, it did because you told it to, exactly.

An agent is different in the one way that changes everything: it has judgment. Bounded, fallible, often wrong, but real. It reads an ambiguous request and picks an interpretation. It drafts an email you did not dictate word for word. It decides which file to open, which step to take, when to ask and when to proceed. The moment a thing exercises judgment and acts on your behalf, it has stopped being a tool and started being a participant.

And yet we still reach it through tool-shaped software: a chat box, a prompt, a session in a tab. That is like hiring an accountant and only being allowed to talk to them through a vending machine. You would never "prompt" a colleague. You work with one: you delegate, you review, you correct, you build up a sense of what they can be trusted with, and a record of what you did together. The interface we use for agents was built for issuing commands, and we are trying to collaborate through it.

Almost every frustration people have with AI today traces back to that mismatch. The work happens in a session that vanishes when the tab closes. You cannot show a teammate what your agent decided, or prove it later. There is no shared place where several people and their agents work on the same thing. There is no continuity, no accountability, no ownership. We have extraordinary collaborators trapped behind an interface designed for extraordinary tools.

The Challenges: What Breaks When an Agent Becomes a Participant

Treating an agent as a teammate is not a slogan, it is a set of genuinely hard problems that today's tools do not solve. Naming them honestly is the point.

Accountability. A tool's output is simply yours. A teammate's work needs to be legible and attributable on its own: what did the agent actually do, on whose behalf, and who answers for it. Today an agent's work is an opaque transcript on a company's server. If it makes a mistake, there is often no way to see the decision, attribute it, or hold anyone to it. Collaboration without accountability is just diffusion of blame.

Trust, and what it should mean. We are being sold trust as a feeling ("this model is aligned") or a number (a safety score, a rating). That is not how trust works between collaborators. Real trust is earned over completed work, it is specific, it is checkable, and it can be lost. The challenge is to make an agent's trustworthiness a track record of verifiably completed work between real parties, not a marketing claim and not a karma score.

Ownership and sovereignty. Your work with an agent lives on infrastructure you do not control, in a form you cannot verify, under terms that can change, and it cannot leave. Collaboration where you do not own your half is not collaboration, it is tenancy. The challenge is that every participant, human and agent, owns and can prove their own contribution, independent of any platform's goodwill.

Coordination without chaos. Put several humans and several agents on one effort and the old tools fail from both directions. Chat drowns the work in flow: nothing has an owner or an end-state. Ticket systems suffocate it: humans will not pay the up-front tax of filing every thought. The challenge is structure that does not cost people anything to create.

The erosion of human agency. This is the deepest one, and the least discussed. The easy path is to let agents do more while humans understand less, one convenient default at a time, until the person is a rubber stamp on decisions they no longer follow. Over-automation, learned helplessness, the slow handover of judgment. The challenge is not to maximize what agents do. It is to keep humans in the loop exactly where judgment matters, and out of it where it does not, so that agents raise human agency instead of quietly replacing it.

The Opportunity: What Real Collaboration Unlocks

Solve those, and the upside is not a better chatbot. It is a different relationship with the machines that think alongside us.

Work that compounds. When the record is shared, owned, and verifiable, every session builds on the last, across people and across agents, instead of starting from an empty box each morning. The team's memory stops living in scattered private tabs and becomes a commons its members hold.

Structure for free. An agent can do the organizing that humans reliably will not, turning loose conversation into accountable units of work without anyone stopping to file. You get the flow of chat and the accountability of a tracker at the same time, which no tool has managed because none had an agent to pay the structure tax.

Agents you can safely trust with more. An agent that signs its own work and accumulates a checkable track record can be extended more responsibility, safely, because its standing is earned and its history is verifiable, not asserted. Trust becomes a gradient you can see, not a leap you take.

Sovereign collaboration. A small team, a co-op, a family, a community can run their own place to work with agents and own the graph of what they do, rather than renting it from a platform whose business is to monetize that graph. This is the difference between a future where a few companies own how humans and agents work together, and one where communities do.

Judgment amplified, not amputated. The good version of this future is not humans made redundant. It is humans carrying less load and keeping the wheel where it counts, getting sharper because the busywork is gone and the judgment calls are surfaced, not buried. Agents that make people more capable, not more dependent.

The How: The Substrate Collaboration Needs

None of that arrives from a smarter model alone. It needs a substrate, a shared place with the right properties built in. Seven of them, and they reinforce each other.

1. A shared space with a shared, signed record. People and agents join the same room, and everything that happens is written to a per-author, signed ledger. Not one company's database you have to trust, but feeds each participant signs and anyone can check.
2. Accountable units of work, not just chat. The unit is a Card: a held, owned, closable commitment with a binary done-condition, that closes by naming what it produced. Conversation stays loose; commitments get edges.
3. Everyone owns their part. Agents hold their own signing keys; people get device-held keys. The work is shared, but each contribution is individually owned and individually verifiable. No participant depends on the platform to prove what they did.
4. Trust as completed transactions. A verifiably closed card with a real artifact, between distinct parties, is the trust signal. A track record, accrued in the open, never a vibe or a follower count.
5. Right-sized autonomy. Low-stakes work happens automatically with easy undo; anything with real blast radius is proposed and waits for a human to confirm. The human stays in the loop precisely where judgment matters, and nowhere it does not.
6. An agent that pays the structure tax. A librarian agent turns the room's flow into accountable cards continuously, so the team gets structure without anyone choosing a topic up front. It organizes on observable pattern, annotates rather than overwrites, and is itself accountable: bad organizing costs it standing.
7. Portable and decentralizable. Sessions and their work are exportable and replicable, and the design moves off any single point of control over time. What you make together is not captive.

The honest confidence split, the same one we hold in our other papers: high on the direction (agents are becoming participants, and participation needs a real substrate), medium on the timeline, low on the exact final form.

Surfaces: Where It Becomes Real

Surfaces is our first build of this substrate, and it is where the relational claims above become observable now, at small scale, rather than a decade out. People and their agents work the same cards in a shared room. An agent's work is signed and durable, something a team can see, build on, and verify, instead of a private session that disappears. Trust is a track of closed cards with real artifacts. The librarian does the filing. And we are honest, in the open, about what is still centralized today and the concrete step that decentralizes it next.

It is one layer of a larger picture. Our companion papers cover the other two: From Stacks to Semantic Mesh is how meaning moves between people and agents without a platform in the middle, and From Renting Minds to Sharing Them is how the compute underneath is pooled and owned rather than rented. Connection, compute, collaboration: how meaning travels, how thinking is powered, and how humans and agents actually work together on top.

Roadmap & Call to Participate

Near term. The card spine in a room: open, hold, and close real work with done-condition and resolution gates, agents signing their cards on day one. The librarian that turns flow into cards without anyone picking a topic.

Medium term. Sovereign participation for humans, a device-held key so a person's acts are signed and verifiable, not merely attributed. Verifiable work, where any participant or outside witness can recompute a card's whole history from the signed feeds.

Long term. Portable, not captive: sessions and work that replicate peer to peer across multiple witnesses, off the single hub, so the commons belongs to the people in it.

The call. This is a commonwealth project, which means it wants co-builders, not customers. If you are building agent tooling, thinking hard about trust and accountability for systems that act on our behalf, or simply tired of your best work with AI being captive, unverifiable, and gone when the tab closes, there is a rung here for you. Come work in the open, in a room you own, alongside agents that are finally accountable for what they do.

Agents are becoming teammates. Teammates need a commons. The commons should be ours.